Type | |
Stats | 69 13 |
Reviews | (14) |
Published | Sep 23, 2024 |
Base Model | |
Training | Steps: 64,563 Epochs: 108 |
Usage Tips | Strength: 1 |
Trigger Words | frisk_(undertale) chara_(undertale) sans_(undertale) toriel_(undertale) asgore_(undertale) mettaton_(undertale) alphys_(undertale) undyne_(undertale) muffet_(undertale) grillby_(undertale) +5 more |
Hash | AutoV2 85F890D61C |
I've noticed there is a lack of Undertale LORAs for SDXL, so I've trained one on Animagine XL v3.x, but can work on some other models (I also trained PonyXL version)
I used around 800 images in total that I could find, with some additions later on. Suffice to say, it can influence style quite a bit (mostly 3.0 models)
And after a long time, Flux LoRA is also being trained. I am not sure about its quality, though, as I am training it by using split mode and fp8 version of dev model.
Characters:
Asriel, Chara, Frisk, Muffet, Sans, Papyrus, Undyne, Gaster, Toriel, Mettaton, Grillby, Alphys, Flowey, Asgore, Napstablook
Some better, some worse. But I don't recommend trying to generate either Asgore or Napstablook in the version 1.0 - they are overfitted. In the version 1.1 and onwards they work well enough, but be careful about it (but it is usually obvious). There could be some stylistic inconsistency, so it is better to specify.
Parameters:
Trigger words are supposed to be like character_(undertale). Some chararacters can be accurate enough with trigger words alone, but for others it is better to add some more.
For Flux it would be a natural language prompting, while SDXL ones are using tagging. Strictly speaking, the _(undertale) part isn't completely necessary as it undestands it without it.
Flux:
It is better to use guidance of 3.5 or higher, otherwise it might lose some of the character likeness, but it would increase the stylistic difference between images.
Animagine 3.1:
Starting with model 1.2, it surpasses the previous models in its flexibility and capabilities. But stylistically they are quite different, so it is a matter of taste. Personally I like 1.2.5 model for its style, but it has lots of biases since I overfitted it a bit in the beginning, which makes it hard to generate some things.
Base XL:
I've trained the second model, but this time on a base model, which is supposed to make it possible to use LoRA on any model, but not really (anime models seem to struggle a lot). So my examples are not only from Animagine v3 on this one. As of now, it isn't wholly accurate, but it really depends on the model. I trained this model as a substitute, if the other one doesn't work at all.
PonyXL:
This version was trained on civitai, while generally it can have good outputs - it missed some of the characters' features.